Elamur ResearchβŒ•
☰
home / papers / 2026 / Whose Pragmatics? Cultural Grounding as a Bottleneck for Stereotype Detection in Egyptian Arabic Social Media

Whose Pragmatics? Cultural Grounding as a Bottleneck for Stereotype Detection in Egyptian Arabic Social Media

Summary

Popular bias benchmarks (StereoSet, CrowS-Pairs, BBQ) test only the lexicon: which words appear next to the names of social groups. The author shows that a stereotype is often conveyed indirectly, through presupposition, implicature, or the form of the speech act. Meanwhile, the surface text remains neutral or even approving. The author calls this phenomenon pragmatic stereotyping. To study it, GPT-4 and Claude 3.5 Sonnet classified a stratified sample of 500 Egyptian Arabic comments from social media. The comments were annotated with a seven-tag scheme that combines sentiment and (im)politeness. To determine at which reasoning step a model fails, the paper proposes SI-CoT, a five-level diagnostic scheme. About 35% of errors are due to a lack of cultural grounding. The models fail to notice bias hidden in "compliments with a subtext." At the same time, they apply English-language pragmatic norms: they read sincere polite criticism as sarcasm and friendly teasing as hostility. For content moderation in languages other than English, the conclusion is as follows: lexical tests are insufficient, and models need cultural-pragmatic competence.

Key points

  • The annotation has seven tags. The models recognize the three "explicit" ones (PP, PoN, IN) well, where form matches content: F1 0.78–0.85. The four "implicit" ones (IP, PN, MPN, MIP) yield an F1 of only 0.62–0.69. On average this is 0.81 versus 0.66, a gap of 15 percentage points.
  • The errors go in both directions. On the comment author's side (the IP tag), models often perceive hidden bias as humorous. On the model's side, Anglocentric notions of politeness distort the reading of PN and MIP.
  • SI-CoT has five levels. The first three rely on surface features. Levels 4–5 (the social goal of the utterance and the politeness assessment) depend on culture, and this is precisely where the failures occur.
  • The scheme was validated: inter-annotator agreement among experts was ΞΊ = 0.78, and it was additionally checked by 200 native speakers. IP comments were correctly classified in about 66% of cases, MIP in 67%. Confident MIP predictions are correct in only 71% of cases.
  • Limitations: only two models, one dialect, and one dataset. Rare tags are scarce in the test sample (IP = 32). Few-shot examples only partially narrow the gap.

Our summary of the paper, in our own words β€” not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.stereacult-1.7

@inproceedings{assem2026whose,
  title = {{Whose Pragmatics? Cultural Grounding as a Bottleneck for Stereotype Detection in Egyptian Arabic Social Media}},
  author = {Assem, Samar A.},
  booktitle = {Proceedings of the 1st Workshop on Stereotypes Across Cultures in Language Technologies (StereACuLT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.stereacult-1.7},
  url = {https://doi.org/10.18653/v1/2026.stereacult-1.7}
}

References

9 references, 3 of them with a card of ours. β†— leads to the work, Β«findΒ» searches for one we have no id for yet.

  1. X. Chen and S. Wang. 2025. Pragmatic inference chain (PIC): Improving LLMs’ reasoning of authentic implicit toxic language. arXiv. β†—
  2. Ibrahim Abu Farha and Walid Magdy. 2020. From Arabic sentiment analysis to sarcasm detection: The ArSarcasm dataset. In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and Processing Tools, pages 32–39. European Language Resource Association. find our card
  3. Y. Cho and S. Kim. 2024. Pragmatic inference of scalar implicature by LLMs. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 10–20. Association for Computational Linguistics. β†— our card
  4. Ibrahim Abu Farha and Walid Magdy. 2021. Benchmarking transformer-based language models for Arabic sentiment and sarcasm detection. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 21–31. Association for Computational Linguistics. find our card
  5. Jonathan Culpeper. 2011. Politeness and impoliteness. In Wolfram Bublitz, Andreas H. Jucker, and Klaus P. Schneider, editors, Pragmatics of Society, volume 5, pages 393– 438. De Gruyter Mouton. find
  6. Q. Abuein, R. M. Al-Khatib, A. Migdady, M. S. Jawarneh, and A. Al-Khateeb. 2024. ArSa-tweets: A novel Arabic sarcasm detection system based on deep learning model. Heliyon, 10(17):e36892. β†—
  7. H. Paul Grice. 1975. Logic and conversation. In Peter Cole and Jerry L. Morgan, editors, Syntax and Semantics 3: Speech Acts, pages 41–58. Academic Press, New York. β†—
  8. R. Mao, M. Ge, S. Han, W. Li, K. He, L. Zhu, and E. Cambria. 2024. A survey on pragmatic processing techniques. Information Fusion, 114:102712. β†—
  9. Mahmoud Al-Ayyoub, Abed Allah Khamaiseh, Yaser Jararweh, and Mohammed N. Al-Kabi. 2019. A comprehensive survey of Ara-A Additional Example Comments R. Mustafin. 2025. Pragmatic perspective on assessing implicit meaning interpretation in sentiment analysis models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 898–907. Association for Computational Linguistics. A.1 Author-Side Pragmatic Stereotyping (IP β†’ MIP misclassification) find

This page in Markdown β†’ https://md.research.elamur.ai/free/papers/assem-2026-whose-pragmatics-cultural-grounding-as/